High-Resolution Image Reconstruction with Unsupervised Learning and Noisy Data Applied to Ion-Beam Dynamics for Particle Accelerators
This work addresses the challenge of high-resolution reconstruction of beam halo structures in high-energy physics accelerators under conditions of strong noise and severe degradation, where conventional methods encounter performance bottlenecks. The authors propose an unsupervised learning framework that requires no training data, integrating convolutional filtering with neural networks and incorporating an optimized early-stopping strategy to mitigate overfitting. This approach enables robust denoising and high-fidelity reconstruction of beam emittance images at low signal-to-noise ratios. Notably, it achieves high-resolution recovery of beam images without ground-truth labels for the first time, extending measurable amplitudes beyond seven standard deviations and significantly enhancing the resolution of beam halo features—thereby overcoming limitations of existing techniques.